Embodied City addresses the gap in embodied intelligence research for complex, open-world urban environments. It provides a benchmark platform, including a high-fidelity simulator and datasets, enabling the development and evaluation of AI agents capable of human-like interaction within cityscapes. This platform is targeted at AI researchers and engineers seeking to advance embodied AI beyond constrained indoor settings, offering a realistic simulation for training and testing navigation and interaction tasks.

How It Works

The simulator leverages Unreal Engine to construct a detailed 3D model of a Beijing business district, incorporating manually created assets (via Blender) for buildings, streets, and urban elements, grounded in real-world data from Baidu Map and Amap. Realistic dynamic elements like vehicles and pedestrians are simulated using the Mirage Simulation System. This approach provides a high degree of visual fidelity and behavioral realism, crucial for training embodied agents in complex urban scenarios. The platform supports Vision-Language Navigation (VLN) tasks, integrating multimodal models to interpret natural language instructions for agent guidance.

Quick Start & Requirements

• Installation: Clone the repository and set up the Python environment using Conda:

  conda env create -n EmbodiedCity -f environment.yml
  conda activate EmbodiedCity
  

  Alternatively, for Python 3.10:

  conda create -n EmbodiedCity python=3.10
  conda activate EmbodiedCity
  pip install -r requirements.txt
  

• Prerequisites: Conda, Python 3.10 (recommended), AirSim, and potentially API keys for online deployment. The offline simulator is noted as compatible with Windows systems.
• Resources: Setup involves environment configuration and potentially downloading large simulator assets.
• Links: Simulator download: [link], Paper: [link].

Highlighted Details

• A benchmark platform for embodied intelligence in real-world city environments.
• High-fidelity 3D urban simulation built on Unreal Engine, featuring detailed buildings, streets, and dynamic agents.
• Supports Vision-Language Navigation (VLN) tasks using multimodal models (OpenAI, Claude).
• Evaluates agent performance using Success Rate (SR), SPL, and Navigation Error (NE) metrics.

Maintenance & Community

Information regarding maintainers, community channels (e.g., Discord, Slack), active development, or roadmap is not detailed in the provided README.

Licensing & Compatibility

The README does not specify a software license, leaving the terms of use, distribution, and modification unclear. Compatibility for commercial use or integration with closed-source systems requires clarification.

Limitations & Caveats

The offline simulator is explicitly stated as compatible with Windows, suggesting potential limitations or lack of support for other operating systems. Access to the online deployed environment requires requesting an API key, which may involve usage restrictions or costs. The focus on urban environments, while detailed, may not encompass all aspects of "open-world" scenarios.